Pith. sign in

REVIEW 3 major objections 4 minor 61 references

This paper shows that a purely label-free pipeline—fakeprint extraction, non-negative matrix factorization, and a Gaussian-blur comparison—can separate real from AI-generated music and cluster different generators, even when the generators

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 02:08 UTC pith:AGF2F2A3

load-bearing objection First zero-shot AI-music detection paper with a clever NMF/blur trick, but the artifact-regularity assumption is contradicted by two of the tested generators — worth serious refereeing, with revisions. the 3 major comments →

arxiv 2607.25530 v1 pith:AGF2F2A3 submitted 2026-07-28 cs.SD

Finding the noise: Zero-shot AI Music Detection

Darius Afchar , Romain Hennequin This is my paper
classification cs.SD
keywords AI music detectionzero-shot learningone-class classificationnon-negative matrix factorizationfakeprintsunsupervised clusteringspectral artifactsaudio forensics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes a zero-shot method for AI music detection where the generating services are unknown or updated. It treats real music as a single known class and uses one-class classification on 'fakeprints'—time-averaged spectral peak profiles—to flag anything outside. The key mechanism: non-negative matrix factorization learns a dictionary of atoms; Gaussian-blurring the atoms and comparing reconstructions yields an error r_i that is low for real music and high for synthetic music. On three datasets covering 18 generators, the binary variant achieves near-perfect detection for most services at a 10% false-positive rate, and the clustering variant (NMF+UMAP+HDBSCAN) separates services, versions, and even engines with high purity. The method is fast, free of deep learning, and does not require any synthetic labels.

Core claim

The central claim is that the delocalized reconstruction error r_i = ||(HW)_i - (H(W*G))_i||_2 is a zero-shot indicator of synthetic content: when NMF factors a matrix of fakeprints, it learns peaked, structured atoms for AI-generated classes and flat, featureless atoms for real music, so blurring the atoms changes the reconstruction of synthetic samples much more than real samples. The paper verifies this on FMA-AE, Echoes, and a new PopularAISet, showing that most classes are detected at rates comparable to supervised baselines, and that the learned clusters correspond to ground-truth generators—including discovery that one service is built on another and that different versions of a model

What carries the argument

The fakeprint is the paper's core representation: for a music track, compute the time-averaged power spectrum from a short-time Fourier transform, subtract local minima to extract residual peaks, restrict to a frequency band, and normalize. Non-negative matrix factorization (NMF) with elastic-net regularization learns a small dictionary of atoms from the fakeprint matrix; the crucial step is comparing the reconstruction with the original atoms to the reconstruction with atoms convolved by a Gaussian kernel. The resulting per-sample error r_i acts as an edge/peak detection score. For clustering (Task B), the same NMF activations are passed through UMAP (a manifold projection) and HDBSCAN (a d

Load-bearing premise

The method assumes that AI-generated music consistently shows localized periodic peaks in its mean power-spectrum fakeprints, while real music shows randomly scattered peaks that average to a flat profile—if a generator's fakeprints lack that consistent periodicity, the NMF reconstruction error will treat its output as real, as the paper's own Mubert and Mureka (pre-v9) results demonstrate.

What would settle it

Generate or collect a large set of tracks from a modern generator that does not use transposed-convolution upsampling—e.g., a diffusion or flow-matching model with anti-artifact post-processing—and run the proposed pipeline. If the distribution of reconstruction errors r_i overlaps the real-music distribution (or the average fakeprint is flat), the claim of zero-shot detection fails for that generator. A cheaper check: measure r_i on Mubert tracks; the paper reports only 5% detection at 10% FPR, so Mubert is already a near-counterexample.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Streaming platforms can use this as a label-free first-pass alarm: any surge of tracks with high r_i from a single label or artist triggers a manual review, without needing to retrain on the new generator.
  • The one-class calibration makes the false-positive rate on real music explicit and controllable by choosing a quantile of real-music errors, such as 1% or 10%.
  • The clustering variant can track generator updates: since Suno v3.5 separates from v4.5/v5, the method can reveal when a service changes its architecture.
  • The method can expose hidden dependencies between services—it clustered Brev with Suno, consistent with Brev using Suno's engine.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The method is fundamentally bounded by the artifact-regularity assumption: if a future generator removes spectral periodicity (e.g., via anti-artifact post-processing or a non-convolutional architecture), the error r_i will classify it as real. A testable extension is to combine fakeprints with other artifact views (phase, time-domain) to widen coverage.
  • The blur-difference criterion is a general 'structured vs unstructured' detector; the same NMF+blur recipe could be applied to other synthetic media (images, speech) that exhibit periodic artifacts, though the overlap with real content would need to be calibrated.
  • The paper's failure on Mubert and Mureka pre-v9 suggests a hidden cost: unsupervised detection works best when the synthetic class is internally consistent and has a large enough sample count; small or heterogeneous synthetic batches may be absorbed into the real class—an operational risk the paper acknowledges.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes a zero-shot pipeline for AI-music detection. It builds on the authors' earlier 'fakeprint' representation (a residual-peak profile of the mean power spectrum) and applies non-negative matrix factorization (NMF) to a mixture of real and synthetic fakeprints. For Task A (real vs. synthetic), it reconstructs each sample from NMF atoms, blurs the atoms with a Gaussian kernel, and uses the per-sample reconstruction error r_i as a one-class score, with a threshold calibrated on real examples to target a false-positive rate. For Task B (multi-class identification), it applies UMAP and HDBSCAN to the NMF activation space. Experiments on FMA-AE, the Echoes dataset, and a newly collected PopularAISet report high detection rates for most generators at 10% FPR, good per-cluster purity in UMAP plots, and the ability to discover version-level structure (e.g., Suno v3.5 vs. later versions). The paper explicitly notes that Mubert and pre-v9 Mureka are not detected by fakeprints.

Significance. If the claims held, the paper would make a useful contribution: a fast, non-deep-learning, zero-shot complement to supervised detectors for monitoring unknown AI-music services. Its strengths include testing on an external dataset (Echoes) and a newly collected set, a supervised separability check as an independent probe of the representation, and the observation that the pipeline can separate model versions. The key limitation is that the method's central mechanism is an empirical regularity about periodic fakeprint peaks, and the paper's own results show that this regularity fails for two of the eighteen generators, so the zero-shot generalization claim must be substantially qualified.

major comments (3)
  1. [§3.1, §3.3, Table 2] The artifact-regularity assumption is load-bearing but contradicted by the paper's own results. The method assumes synthetic fakeprints have consistent localized periodic peaks while real fakeprints average flat, so NMF learns peaked atoms for synthetic classes and flat atoms for real music. Table 2 shows Mubert detection at 5% at 10% FPR and 0% at 1% FPR, and Mureka at 48.5%/14.9%; Figure 2 places Mubert and pre-v9 Mureka in real-mixed clusters. Thus the method detects a specific artifact class, not AI music in general. The claims of 'excellent performance' and suitability for monitoring 'various newly released generative models' are too strong. The paper should either reframe the contribution as detecting the specific artifact pattern, or provide an analysis that distinguishes generators with and without this pattern and discusses how the detector behaves in the latter case.
  2. [§4.4, Task B, Figure 2] Task B is evaluated only through UMAP/HDBSCAN visualizations and per-cluster purity. There is no aggregate clustering metric such as ARI, NMI, or the number of correctly discovered clusters, nor any comparison to a baseline. The textual claim that classes are 'almost perfectly separated' is not quantitatively supported. The authors should report cluster-level and overall metrics (e.g., adjusted Rand index, normalized mutual information, mean purity weighted by cluster size), ideally with confidence intervals given the small Echoes per-class sizes. This is necessary for the zero-shot clustering claim.
  3. [§3.3, §4.4, code repository] The method relies on several free parameters — NMF component count f, Gaussian blur radius, UMAP n_neighbors, HDBSCAN min_cluster_size, elastic-net weights, and fakeprint STFT parameters. The paper states that 'same settings' are used for all datasets but does not report these settings beyond a few values, and the code repository is promised only 'after acceptance.' More importantly, there is no sensitivity analysis showing how r_i and the resulting detections depend on f, the blur radius, or the regularization weights. Because the core criterion r_i is defined by a specific blur operation, the paper should include an ablation or at least a robustness study demonstrating that the results are not a knife-edge artifact of the chosen hyperparameters.
minor comments (4)
  1. [§4.1, Echoes] The Echoes dataset is very small, with per-class test sets of roughly 30–60 samples after splitting. The paper acknowledges this but Table 1 and Table 2 present point estimates with no error bars or statistical tests. Please add confidence intervals or bootstrap estimates.
  2. [Figure 2] The UMAP plots with ellipses and purity values are hard to read in grayscale and the color legend is not fully described. A tabular summary of cluster composition and purity would improve the presentation.
  3. [§4.2] Typo: 'an_fft' should be 'n_fft'. Also, the sentence 'We use the same set of settings' is awkward; recommend 'We use the same hyperparameter configuration'.
  4. [§5/Conclusion] The conclusion honestly states the limitation about sample size and silent treatment of generators without periodic structure. This is appropriate, but it should be moved to or repeated in the abstract and Section 1, because it substantially narrows the zero-shot claim.

Circularity Check

0 steps flagged

No significant circularity; fakeprint assumption is empirically validated rather than constructionally forced.

full rationale

The paper's derivation chain is: fakeprint(x) from prior work [17] -> NMF dictionary learning (H, W) -> Gaussian-blur reconstruction error r_i -> threshold calibrated on real-only samples. No step reduces a 'prediction' to a fitted input. In Task A, synthetic labels are never used: the threshold is chosen as a quantile of errors on real examples only (Sec. 3.3: 'given a small training collection of real examples, we take a quantile of a training set of errors...'), and the reported synthetic detection rates are then measured on unseen synthetic classes. The NMF factorization is learned on the full fakeprint matrix, including test samples, which is transductive but not circular. The core artifact-regularity assumption is inherited from the same authors' [16,17], but it is not relied on by citation alone: Sec. 4.3 explicitly tests linear separability of fakeprints on 18 models, including the external Echoes dataset, and Figure 1 provides a direct visualization of the peak-vs-flat distinction. The paper's own negative results for Mubert and pre-v9 Mureka are acknowledged limitations (Conclusion: 'We should also further investigate why samples from Mureka (before v9) and Mubert do not seem to be detected by the fakeprints'), which actually indicate the criterion is not trivially forced by construction. Task B uses UMAP/HDBSCAN on unsupervised NMF activations with hyperparameters set from an explicit class-count upper-bound assumption, not from synthetic labels. There is no equation where the output equals an input by definition, no fitted parameter renamed as a prediction, and no self-citation chain that carries the central claim alone. The only self-citation is the use of the authors' own fakeprint representation; since it is independently validated in this paper, it is not load-bearing.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The method's load-bearing ingredients are two domain assumptions about spectral artifacts in AI music (from the authors' prior work) plus several hand-set hyperparameters; none of the free parameters are fitted to the target labels, but f is chosen using knowledge of the number of classes.

free parameters (6)
  • NMF component count f = 20 (fixed across datasets)
    Set as an upper bound on the number of classes; the paper knows the true class counts (4+real to 10+real), so this leaks dataset knowledge into a zero-shot method (Sec. 4.2).
  • Gaussian kernel radius for blur = 10
    Chosen as 'quite large' relative to peaks; controls the reconstruction-error criterion r_i; no sensitivity analysis (Sec. 4.2).
  • UMAP n_neighbors = 20
    'Arbitrarily fixed at 20'; affects Task B cluster geometry (Sec. 4.2).
  • HDBSCAN min_cluster_size = n/5f
    'Arbitrary fraction of the average expected cluster size'; affects number and purity of clusters (Sec. 4.2).
  • Elastic-net regularization weights lambda_H, lambda_W = not specified
    Values deferred to a not-yet-public code repository; needed to reproduce NMF dictionary (Sec. 3.3).
  • Fakeprint STFT parameters = n_fft=2^14; band 3-15 kHz; vector length 4458
    Inherited from prior work [17] with minor adjustments; the representation itself is the main input feature and is not re-derived here (Sec. 4.2).
axioms (5)
  • domain assumption Deconvolution layers of CNNs leave periodic energy peaks ('checkerboard artifacts') in generated audio, and fakeprints capture these.
    Borrowed from [17] by the same group; it is the foundation for the whole method; verified in Sec. 4.3 for most but not all generators (Mubert, Mureka fail).
  • domain assumption Real music has no consistent spectral peak pattern in fakeprints; peaks are randomly located and average to flat.
    Stated in Sec. 3.1 and Figure 1; if false, the NMF denoising and blur criterion cannot separate real from synthetic.
  • ad hoc to paper NMF denoising will learn peaked atoms only for classes that appear in sufficient number; small synthetic classes will be absorbed into flat atoms.
    Used to justify why small sample collections or weakly-artifacted classes (Mubert, Mureka pre-v9) are treated as real; Sec. 3.3 and Conclusion acknowledge this.
  • ad hoc to paper Gaussian blur destroys peaked atoms but leaves flat atoms unchanged, making reconstruction error r_i a valid real/synthetic criterion.
    Core heuristic in Sec. 3.3; no derivation or theorem; tested only indirectly via results.
  • ad hoc to paper HDBSCAN and UMAP parameter choices produce meaningful clusters aligned with generator identity.
    Defaults are arbitrary fractions (Sec. 4.2); no sensitivity analysis or quantitative cluster metric.

pith-pipeline@v1.3.0-alltime-deepseek · 11703 in / 13842 out tokens · 125557 ms · 2026-08-01T02:08:47.102454+00:00 · methodology

0 comments
read the original abstract

We present a novel method for AI-generated music detection in scenarios where the models that generated the input samples are unknown to the detector (e.g., from a newly released service). Since 2023, there has been a multiplication of user-friendly AI-music generation services (e.g., Suno, Udio), along with regular updates and new features. There is thus a need to address synthetic content detection in an unsupervised way to adapt to this rapidly changing context. This angle has not been much studied in music yet. We propose to study two tasks. First, discriminating between real and synthetic music. This may be approached in a one-class manner, namely, using some baseline real music and trying to determine what falls outside. Second, zero-shot multi-class identification, which is more similar to an unsupervised clustering task on a mix of real and various AI-music generations, where the goal is to create coherent, high-purity clusters. We propose a combination of a previously proposed artifact-extraction method, on top of which we apply non-negative matrix factorization and simple classification and clustering methods. We achieve excellent performance on both tasks, showing that the proposed methods may be used to monitor large-scale catalogs that may receive AI-generated samples from various newly released generative models.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

61 extracted references · 8 linked inside Pith

  1. [1]

    fair use

    INTRODUCTION Since the end of 2023, there has been a massification of AI- generated music with the launch of services such asSuno andUdio. These services follow the paradigm of popu- lar tech services like OpenAI’s ChatGPT in that they pro- pose to fully generate music in a matter of seconds from a text prompt, and are available online for free. Since the...

    2023

  2. [2]

    AI Slop invasion

    RELA TED WORK Currently, there are relatively few work on the topic of AI- music detection [13, 15–21]. The topic is quite novel: the first papers were published in 2024, following the release arXiv:2607.25530v1 [cs.SD] 28 Jul 2026 of Suno at the end of 2023 which dramatically increased the volume of AI-generated music uploaded to the inter- net [1,2,22–2...

    Pith/arXiv arXiv 2024

  3. [3]

    checkerboard artifacts

    METHOD We provide some background on AI-music artifacts, for- malize the tasks we propose to study, and present our method to achieve such an aim. 3.1 Background To understand the reasonings behind our method, we must briefly introduce the AI-music artifact extraction process proposed in [17]. Simply put, the deconvolution layers of convolutional neural n...

  4. [4]

    We only want an atom to be learned if a sufficient collection of fakeprints follow a given pattern

    This enables to have a well- behaved dictionary of atoms. We only want an atom to be learned if a sufficient collection of fakeprints follow a given pattern. Conversely, we do not want each unique fakeprint of real music to be learned. Said differently,real music is treated as random noisein the fakeprint represen- tation. NMF has been used in audio with ...

  5. [5]

    Encodec, DAC, Musica, GrifMel

    EXPERIMENTS We present our datasets, some preliminary supervised re- sults to check if the data is separable using the fakeprints, and the experiments for our two tasks and method. 4.1 Datasets As exposed in Section 2, there are relatively few work published on AI music detection, which also means few published datasets (cf.the survey [54]). For our tasks...

  6. [6]

    CONCLUSION We show that previous work conducted on AI music de- tection can be well extended to the zero-shot territory. Our method is fast, does not use deep learning, and can serve as a complement to existing supervised detectors to track newly released versions and models, paving the way for the exhaustive monitoring of AI-generated music in deliv- eri...

  7. [7]

    Deezer deploys cutting-edge AI detection tool for music streaming,

    Deezer, “Deezer deploys cutting-edge AI detection tool for music streaming,” https://newsroom-deezer. com/2025/04/deezer-reveals-18-of-all-new-music-u ploaded-to-streaming-is-fully-ai-generated/, 2025, [Online; accessed 21-April-2026]

    2025

  8. [8]

    Deezer deploys cutting-edge AI detection tool for music streaming,

    ——, “Deezer deploys cutting-edge AI detection tool for music streaming,” https://newsroom-deezer.com/ 2026/04/ai-generated-tracks-represent-44-of-new -uploaded-music/, 2026, [Online; accessed 21-April- 2026]

    2026

  9. [9]

    AI Art and its Impact on Artists,

    H. H. Jiang, L. Brown, J. Cheng, M. Khan, A. Gupta, D. Workman, A. Hanna, J. Flowers, and T. Gebru, “AI Art and its Impact on Artists,” inAIES. ACM, 2023

    2023

  10. [10]

    AI Art is Theft: Labour, Extraction, and Exploitation: Or, On the Dangers of Stochastic Pol- locks,

    T. S. Goetze, “AI Art is Theft: Labour, Extraction, and Exploitation: Or, On the Dangers of Stochastic Pol- locks,” inACM FAccT, 2024

    2024

  11. [11]

    Exploring the Use of Abusive Generative AI Models on Civitai,

    Y . Wei, Y . Zhu, P. Hui, and G. Tyson, “Exploring the Use of Abusive Generative AI Models on Civitai,” in Proceedings of the 32nd ACM International Confer- ence on Multimedia, 2024, pp. 6949–6958

    2024

  12. [12]

    Deezer confirms demonetization of up to 85% of AI-music streams due to fraud and moves to sell AI- detection Technology,

    Deezer, “Deezer confirms demonetization of up to 85% of AI-music streams due to fraud and moves to sell AI- detection Technology,” https://newsroom-deezer.com/ 2026/01/ai-generated-music-deezer-selling-detection -tool/, 2026, [Online; accessed 21-April-2026]

    2026

  13. [13]

    EU AI Act,

    European Comission, “EU AI Act,” https://artificialint elligenceact.eu/article/50/, 2024, Article 50

    2024

  14. [14]

    Comparing apples to oranges: A taxonomy for nav- igating the global landscape of AI regulation,

    S. Alanoca, S. Gur-Arieh, T. Zick, and K. Klyman, “Comparing apples to oranges: A taxonomy for nav- igating the global landscape of AI regulation,” inPro- ceedings of the 2025 ACM Conference on Fairness, Ac- countability, and Transparency, 2025, pp. 914–937

    2025

  15. [15]

    Provocations from the Humanities for Generative AI Research,

    L. Klein, M. Martin, A. Brock, M. Antoniak, M. Walsh, J. M. Johnson, L. Tilton, and D. Mimno, “Provocations from the Humanities for Generative AI Research,”arXiv preprint arXiv:2502.19190, 2025

    arXiv 2025

  16. [16]

    Pelly,Mood Machine: The Rise of Spotify and the Costs of the Perfect Playlist

    L. Pelly,Mood Machine: The Rise of Spotify and the Costs of the Perfect Playlist. Hodder & Stoughton, 2025

    2025

  17. [17]

    Study on the economic impact of Generative AI in the Music and Audiovisual industries,

    CISAC, “Study on the economic impact of Generative AI in the Music and Audiovisual industries,” https:// www.cisac.org/Newsroom/news-releases/global-eco nomic-study-shows-human-creators-future-risk-gen erative-ai, 2024, [Online; accessed 22-March-2025]

    2024

  18. [18]

    " i made this (sort of)

    B. L. Sturm, “" i made this (sort of)": Negotiating au- thorship, confronting fraudulence, and exploring new musical spaces with prompt-based ai music genera- tion,”arXiv preprint arXiv:2507.23365, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    SONICS: Synthetic Or Not - Identi- fying Counterfeit Songs,

    M. A. Rahman, Z. I. A. Hakim, N. H. Sarker, B. Paul, and S. A. Fattah, “SONICS: Synthetic Or Not - Identi- fying Counterfeit Songs,” inInternational Conference on Learning Representations (ICLR), 2025

    2025

  20. [20]

    Music Information Re- trieval Evaluation eXchange (MIREX),

    J. Jiang, G. Xia, A. Maezawa, Z. Wang, Y . Zhang, R. Yuan, and J. S. Downie, “Music Information Re- trieval Evaluation eXchange (MIREX),” https://musi c-ir.org/mirex/wiki/2025:Main_Page, 2025, [Online; accessed 27-April-2026]

    2025

  21. [21]

    Singfake: Singing voice deepfake detection,

    Y . Zang, Y . Zhang, M. Heydari, and Z. Duan, “Singfake: Singing voice deepfake detection,” in ICASSP. IEEE, 2024

    2024

  22. [22]

    AI-Generated Music Detection and its Challenges,

    D. Afchar, G. Meseguer-Brocal, and R. Hennequin, “AI-Generated Music Detection and its Challenges,” inICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  23. [23]

    A Fourier Explanation of AI-music Ar- tifacts,

    D. Afchar, G. Meseguer-Brocal, K. Akesbi, and R. Hennequin, “A Fourier Explanation of AI-music Ar- tifacts,” inIsmir 2025 Hybrid Conference, 2025

    2025

  24. [24]

    From Audio Deepfake Detection to AI-Generated Music Detection–A Pathway and Overview,

    Y . Li, M. Milling, L. Specia, and B. W. Schuller, “From Audio Deepfake Detection to AI-Generated Music Detection–A Pathway and Overview,”arXiv preprint arXiv:2412.00571, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    The AI Music Arms Race: On the Detection of AI- Generated Music,

    L. Cros Vila, B. Sturm, L. Casini, and D. Dalmazzo, “The AI Music Arms Race: On the Detection of AI- Generated Music,”Transactions of the International Society for Music Information Retrieval, vol. 8, no. 1, pp. 179–194, 2025

    2025

  26. [26]

    Double entendre: Robust audio-based AI-generated lyrics detection via multi-view fusion,

    M. Frohmann, G. Meseguer-Brocal, M. Schedl, and E. V . Epure, “Double entendre: Robust audio-based AI-generated lyrics detection via multi-view fusion,” inFindings of the Association for Computational Lin- guistics: ACL 2025, 2025, pp. 1914–1926

    2025

  27. [27]

    Echoes: A semantically-aligned music deepfake de- tection dataset,

    O. Pascu, D. Oneata, H. Cucu, and N. M. Muller, “Echoes: A semantically-aligned music deepfake de- tection dataset,”arXiv preprint arXiv:2603.23667, 2026

    Pith/arXiv arXiv 2026

  28. [28]

    Nothing Is Sacred: AI Generated Slop Has Come for Christmas Music,

    Samantha Cole—404 Media, “Nothing Is Sacred: AI Generated Slop Has Come for Christmas Music,” https: //www.404media.co/ai-generated-christmas-music/, 2024, [Online; accessed 25-April-2026]

    2024

  29. [29]

    Musicians are deeply concerned about AI. So why are the major labels em- bracing it?

    Eamonn Forde—The Guardian, “Musicians are deeply concerned about AI. So why are the major labels em- bracing it?” https://www.theguardian.com/music/20 25/dec/16/musicians-are-deeply-concerned-about-a i-so-why-are-the-major-labels-embracing-it, 2025, [Online; accessed 25-April-2026]

    2025

  30. [30]

    AI Slop Is Flooding Streaming—and Musicians Are Fighting Back,

    Andrew R. Chow—Time, “AI Slop Is Flooding Streaming—and Musicians Are Fighting Back,” https: //time.com/article/2026/03/26/ai-slop-is-threateni ng-musicians-can-tech-companies-stem-the-tide-/, 2026, [Online; accessed 25-April-2026]

    2026

  31. [31]

    Blind signal de- compositions for automatic transcription of polyphonic music: NMF and K-SVD on the benchmark,

    N. Bertin, R. Badeau, and G. Richard, “Blind signal de- compositions for automatic transcription of polyphonic music: NMF and K-SVD on the benchmark,” in2007 IEEE International Conference on Acoustics, Speech and Signal Processing-ICASSP’07, vol. 1. IEEE, 2007, pp. I–65

    2007

  32. [32]

    Speech denoising using nonnegative matrix factoriza- tion with priors,

    K. W. Wilson, B. Raj, P. Smaragdis, and A. Divakaran, “Speech denoising using nonnegative matrix factoriza- tion with priors,” in2008 ieee international conference on acoustics, speech and signal processing. IEEE, 2008, pp. 4029–4032

    2008

  33. [33]

    Single-channel audio source separation with NMF: divergences, con- straints and algorithms,

    C. Févotte, E. Vincent, and A. Ozerov, “Single-channel audio source separation with NMF: divergences, con- straints and algorithms,”Audio Source Separation, pp. 1–24, 2018

    2018

  34. [34]

    Tackling interpretability in audio classification networks with non-negative matrix fac- torization,

    J. Parekh, S. Parekh, P. Mozharovskyi, G. Richard, and F. d’Alché Buc, “Tackling interpretability in audio classification networks with non-negative matrix fac- torization,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 1392–1405, 2024

    2024

  35. [35]

    AI slop,

    Wikipedia contributors, “AI slop,” 2026, [Online; accessed 22-April-2026]. [Online]. Available: https: //en.wikipedia.org/w/index.php?title=AI_slop&oldid =1350499922

    2026

  36. [36]

    Opening musical creativity? Embedded ideologies in generative-AI music systems,

    L. Pram and F. Morreale, “Opening musical creativity? Embedded ideologies in generative-AI music systems,” arXiv preprint arXiv:2508.08805, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    Reductive, exclusionary, normalising: the limits of generative AI music,

    F. Morreale, M. A. Martinez-Ramirez, R. Masu, W. Liao, and Y . Mitsufuji, “Reductive, exclusionary, normalising: the limits of generative AI music,”Trans- actions of the International Society for Music Informa- tion Retrieval, vol. 8, no. 1, 2025

    2025

  38. [38]

    Where does the buck stop? Ethical and political issues with AI in music creation,

    F. Morreale, “Where does the buck stop? Ethical and political issues with AI in music creation,”Transac- tions of the International Society for Music Informa- tion Retrieval, vol. 4, no. 1, 2021

    2021

  39. [39]

    Data Col- lection in Music Generation Training Sets: A Critical Analysis

    F. Morreale, M. Sharma, I.-C. Weiet al., “Data Col- lection in Music Generation Training Sets: A Critical Analysis.” inISMIR, 2023, pp. 37–46

    2023

  40. [40]

    The unwitting labourer: ex- tracting humanness in AI training,

    F. Morreale, E. Bahmanteymouri, B. Burmester, A. Chen, and M. Thorp, “The unwitting labourer: ex- tracting humanness in AI training,”AI & SOCIETY, vol. 39, no. 5, pp. 2389–2399, 2024

    2024

  41. [41]

    Human subsumption in training datasets for music generation,

    F. Morreale, “Human subsumption in training datasets for music generation,” inThe Inner World of Artificial Intelligence. CRC Press, 2026, pp. 13–31

    2026

  42. [42]

    Crawford,The atlas of AI: Power, politics, and the planetary costs of artificial intelligence

    K. Crawford,The atlas of AI: Power, politics, and the planetary costs of artificial intelligence. Yale Univer- sity Press, 2021

    2021

  43. [43]

    Narayanan, , and S

    A. Narayanan, , and S. Kapoor,AI Snake Oil: What Ar- tificial Intelligence Can Do, What It Can’t, and How to Tell the Difference. Princeton University Press, 2024

    2024

  44. [44]

    Slopa- ganda: The interaction between propaganda and gener- ative AI,

    M. Klincewicz, M. Alfano, and A. E. Fard, “Slopa- ganda: The interaction between propaganda and gener- ative AI,”Filosofiska Notiser, vol. 12, no. 1, pp. 135– 162, 2025

    2025

  45. [45]

    J. S. Carbonell,Un taylorisme augmenté. Critique de l’intelligence artificielle. Amsterdam éditions, 2025

    2025

  46. [46]

    Technofascism: AI, Big Tech, and the new authoritarianism,

    M. Coeckelbergh, “Technofascism: AI, Big Tech, and the new authoritarianism,”AI & SOCIETY, pp. 1–14, 2026

    2026

  47. [47]

    From Melting Pots to Misrepresentations: Exploring Harms in Gen- erative AI,

    S. Gautam, P. N. Venkit, and S. Ghosh, “From Melting Pots to Misrepresentations: Exploring Harms in Gen- erative AI,” inGenAICHI ’24 in CHI Conference on Human Factors in Computing Systems, 2024

    2024

  48. [48]

    Genera- tive AI and the Future of the Digital Commons: Five Open Questions and Knowledge Gaps,

    A. Noroozian, L. Aldana, M. Arisi, H. Asghari, R. Avila, P. G. Bizzaro, R. Chandrasekhar, C. Con- sonni, D. De Angelis, F. De Chiaraet al., “Genera- tive AI and the Future of the Digital Commons: Five Open Questions and Knowledge Gaps,”arXiv preprint arXiv:2508.06470, 2025

    Pith/arXiv arXiv 2025

  49. [49]

    Deconvolution and Checkerboard Artifacts,

    A. Odena, V . Dumoulin, and C. Olah, “Deconvolution and Checkerboard Artifacts,”Distill, 2016. [Online]. Available: http://distill.pub/2016/deconv-checkerboar d

    2016

  50. [50]

    Importance of semantic representation: Dataless classification

    M.-W. Chang, L.-A. Ratinov, D. Roth, and V . Sriku- mar, “Importance of semantic representation: Dataless classification.” inAAAI, vol. 2, 2008, pp. 830–835

    2008

  51. [51]

    Zero-data learning of new tasks

    H. Larochelle, D. Erhan, and Y . Bengio, “Zero-data learning of new tasks.” inAAAI, vol. 1, no. 2, 2008, p. 3

    2008

  52. [52]

    Zero-shot learning with semantic output codes,

    M. Palatucci, D. Pomerleau, G. E. Hinton, and T. M. Mitchell, “Zero-shot learning with semantic output codes,”Advances in neural information processing sys- tems, vol. 22, 2009

    2009

  53. [53]

    An exper- imental comparison of one-class classification meth- ods,

    D. de Ridder, D. M. Tax, and R. P. Duin, “An exper- imental comparison of one-class classification meth- ods,” inProc. ASCI’98, 4th Annual Conf. of the Ad- vanced School for Computing and Imaging, 1998, pp. 213–218

    1998

  54. [54]

    Estimating the support of a high-dimensional distribution,

    B. Schölkopf, J. C. Platt, J. Shawe-Taylor, A. J. Smola, and R. C. Williamson, “Estimating the support of a high-dimensional distribution,”Neural computation, vol. 13, no. 7, pp. 1443–1471, 2001

    2001

  55. [55]

    On simple one-class classification methods,

    Z. Noumir, P. Honeine, and C. Richard, “On simple one-class classification methods,” in2012 IEEE Inter- national Symposium on Information Theory Proceed- ings. IEEE, 2012, pp. 2022–2026

    2012

  56. [56]

    Algorithms for non-negative matrix factorization,

    D. Lee and H. S. Seung, “Algorithms for non-negative matrix factorization,”Advances in neural information processing systems, vol. 13, 2000

    2000

  57. [57]

    Fast local algorithms for large scale nonnegative matrix and tensor fac- torizations,

    A. Cichocki and A.-H. Phan, “Fast local algorithms for large scale nonnegative matrix and tensor fac- torizations,”IEICE transactions on fundamentals of electronics, communications and computer sciences, vol. 92, no. 3, pp. 708–721, 2009

    2009

  58. [58]

    UMAP: Uniform Manifold Approximation and Pro- jection,

    L. McInnes, J. Healy, N. Saul, and L. Grossberger, “UMAP: Uniform Manifold Approximation and Pro- jection,”The Journal of Open Source Software, vol. 3, no. 29, p. 861, 2018

    2018

  59. [59]

    Density- based clustering based on hierarchical density esti- mates,

    R. J. Campello, D. Moulavi, and J. Sander, “Density- based clustering based on hierarchical density esti- mates,” inPacific-Asia conference on knowledge dis- covery and data mining. Springer, 2013, pp. 160–172

    2013

  60. [60]

    Detecting Machine-Generated Music with Explainability–A Challenge and Early Benchmarks,

    Y . Li, Q. Sun, H. Li, L. Specia, and B. W. Schuller, “Detecting Machine-Generated Music with Explainability–A Challenge and Early Benchmarks,” arXiv preprint arXiv:2412.13421, 2024

    Pith/arXiv arXiv 2024

  61. [61]

    FMA: A dataset for music analysis,

    M. Defferrard, K. Benzi, P. Vandergheynst, and X. Bresson, “FMA: A dataset for music analysis,” in ISMIR, 2017. [Online]. Available: https://arxiv.org/ab s/1612.01840

    Pith/arXiv arXiv 2017